| Title |
Portrait Shadow Removal using Relighting and Intrinsic Component with Facial shadows |
| Authors |
김수정(Sujeong Kim) ; 이상철(Sang-chul Lee) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.11.69 |
| Keywords |
Deep learning; CNN; Encoder-deocoder network; Shadow removal; Relighting |
| Abstract |
Shadows in portrait photography affect the quality of the photo. Shadows distort the face's structure or color, reducing performance in face recognition and detection tasks. To solve this problem, we propose a method to remove shadows from a single portrait image. We aim to remove the shadows lighting from any direction. Inspired by a relighting method that generates a new image illuminating from the original lighting direction to a different desired lighting direction, the shadow is removed by generating an image illuminating the front light. A ratio image is generated through the encoder-decoder structure and multiplied by the input image to generate a shadow-free image. The normal decoder is added to reflect the facial geometry. It helps generate a shadow-free image by estimating the normal map and the shadow mask using the lighting direction. Since the generated shadow-free image has the unnaturalness of the bright area due to the original lighting, we employ the refine network using the feature masking mechanism to remove it. Our network outperforms state-of-the-art shadow removal methods and achieves performance improvements of 0.04, 2.88, and 0.06 in RMSE, PSNR, and SSIM in the YaleB dataset. |